inspect-ai 0.3.103__py3-none-any.whl → 0.3.105__py3-none-any.whl

inspect_ai/model/_providers/anthropic.py

@@ -22,6 +22,8 @@ from anthropic.types import (
     MessageParam,
     RedactedThinkingBlock,
     RedactedThinkingBlockParam,
+    ServerToolUseBlock,
+    ServerToolUseBlockParam,
     TextBlock,
     TextBlockParam,
     ThinkingBlock,
@@ -31,11 +33,15 @@ from anthropic.types import (
     ToolTextEditor20250124Param,
     ToolUseBlock,
     ToolUseBlockParam,
+    WebSearchTool20250305Param,
+    WebSearchToolResultBlock,
+    WebSearchToolResultBlockParam,
     message_create_params,
 )
 from anthropic.types.beta import (
     BetaToolComputerUse20250124Param,
     BetaToolTextEditor20241022Param,
+    BetaToolTextEditor20250429Param,
 )
 from pydantic import JsonValue
 from typing_extensions import override
@@ -43,6 +49,7 @@ from typing_extensions import override
 from inspect_ai._util.constants import BASE_64_DATA_REMOVED, NO_CONTENT
 from inspect_ai._util.content import (
     Content,
+    ContentData,
     ContentImage,
     ContentReasoning,
     ContentText,
@@ -61,6 +68,10 @@ from .._generate_config import GenerateConfig
 from .._model import ModelAPI
 from .._model_call import ModelCall
 from .._model_output import ChatCompletionChoice, ModelOutput, ModelUsage, StopReason
+from .._providers._anthropic_citations import (
+    to_anthropic_citation,
+    to_inspect_citation,
+)
 from .util import environment_prerequisite_error, model_base_url
 from .util.hooks import HttpxHooks
 
@@ -70,6 +81,14 @@ ANTHROPIC_API_KEY = "ANTHROPIC_API_KEY"
 
 INTERNAL_COMPUTER_TOOL_NAME = "computer"
 
+WEB_SEARCH_COMPATIBLE_MODELS = [
+    "claude-opus-4-20250514",
+    "claude-sonnet-4-20250514",
+    "claude-3-7-sonnet-20250219",
+    "claude-3-5-sonnet-latest",
+    "claude-3-5-haiku-latest",
+]
+
 
 class AnthropicAPI(ModelAPI):
     def __init__(
@@ -232,27 +251,19 @@ class AnthropicAPI(ModelAPI):
             if self.extra_body is not None:
                 request["extra_body"] = self.extra_body
 
-            # make request (unless overrideen, stream if we are using reasoning)
+            # make request (unless overridden, stream if we are using reasoning)
             streaming = (
                 self.is_using_thinking(config)
                 if self.streaming == "auto"
                 else self.streaming
             )
-            if streaming:
-                async with self.client.messages.stream(**request) as stream:
-                    message = await stream.get_final_message()
-            else:
-                message = await self.client.messages.create(**request, stream=False)
 
-            # set response for ModelCall
-            response = message.model_dump()
-
-            # extract output
-            output = await model_output_from_message(
-                self.client, self.service_model_name(), message, tools
+            message, output = await self._perform_request_and_continuations(
+                request, streaming, tools
             )
 
-            # return output and call
+            response = message.model_dump()
+
             return output, model_call()
 
         except BadRequestError as ex:
@@ -269,6 +280,50 @@ class AnthropicAPI(ModelAPI):
             else:
                 raise ex
 
+    async def _perform_request_and_continuations(
+        self,
+        request: dict[str, Any],
+        streaming: bool,
+        tools: list[ToolInfo],
+    ) -> tuple[Message, ModelOutput]:
+        """
+        This helper function is split out so that it can be easily call itself recursively in cases where the model requires a continuation
+
+        It considers the result from the initial request the "head" and the result
+        from the continuation the "tail".
+        """
+        if streaming:
+            async with self.client.messages.stream(**request) as stream:
+                head_message = await stream.get_final_message()
+        else:
+            head_message = await self.client.messages.create(**request, stream=False)
+
+        head_model_output, continuation_required = await model_output_from_message(
+            self.client, self.service_model_name(), head_message, tools
+        )
+
+        if continuation_required:
+            tail_request = dict(request)
+            tail_request["messages"] = request["messages"] + [
+                MessageParam(role=head_message.role, content=head_message.content)
+            ]
+            _, tail_model_output = await self._perform_request_and_continuations(
+                tail_request, streaming, tools
+            )
+
+            head_content = _content_list(head_model_output.message.content)
+            tail_content = _content_list(tail_model_output.message.content)
+            tail_model_output.message.content = head_content + tail_content
+
+            # TODO:
+            # It looks weird to return the head message with the tail output, but
+            # the contract for this function is that it returns the head message
+            # even when it has needed to recurse. This is because model_call()
+            # above doesn't currently support multiple requests
+            return head_message, tail_model_output
+
+        return head_message, head_model_output
+
     def completion_config(
         self, config: GenerateConfig
     ) -> tuple[dict[str, Any], dict[str, str], list[str]]:
@@ -343,6 +398,9 @@ class AnthropicAPI(ModelAPI):
     def is_claude_3_7(self) -> bool:
         return "claude-3-7-" in self.service_model_name()
 
+    def is_claude_4(self) -> bool:
+        return re.search(r"claude-4-[a-zA-Z]", self.service_model_name()) is not None
+
     @override
     def connection_key(self) -> str:
         return str(self.api_key)
@@ -521,7 +579,11 @@ class AnthropicAPI(ModelAPI):
         self, tool: ToolInfo, config: GenerateConfig
     ) -> Optional["ToolParamDef"]:
         return (
-            (self.computer_use_tool_param(tool) or self.text_editor_tool_param(tool))
+            (
+                self.computer_use_tool_param(tool)
+                or self.text_editor_tool_param(tool)
+                or self.web_search_tool_param(tool)
+            )
             if config.internal_tools is not False
             else None
         )
@@ -569,7 +631,17 @@ class AnthropicAPI(ModelAPI):
 
     def text_editor_tool_param(
         self, tool: ToolInfo
-    ) -> ToolTextEditor20250124Param | BetaToolTextEditor20241022Param | None:
+    ) -> (
+        ToolTextEditor20250124Param
+        | BetaToolTextEditor20241022Param
+        | BetaToolTextEditor20250429Param
+        | None
+    ):
+        # See: https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/text-editor-tool#before-using-the-text-editor-tool
+        # TODO: It would be great to enhance our `is_claude_xxx` functions to help here.
+        if self.model_name.startswith(("claude-3-5-haiku", "claude-3-opus")):
+            return None
+
         # check for compatible 'text editor' tool
         if tool.name == "text_editor" and (
             sorted(tool.parameters.properties.keys())
@@ -586,7 +658,11 @@ class AnthropicAPI(ModelAPI):
             )
         ):
             return (
-                BetaToolTextEditor20241022Param(
+                BetaToolTextEditor20250429Param(
+                    type="text_editor_20250429", name="str_replace_based_edit_tool"
+                )
+                if self.is_claude_4()
+                else BetaToolTextEditor20241022Param(
                     type="text_editor_20241022", name="str_replace_editor"
                 )
                 if self.is_claude_3_5()
@@ -598,6 +674,49 @@ class AnthropicAPI(ModelAPI):
         else:
             return None
 
+    def web_search_tool_param(
+        self, tool: ToolInfo
+    ) -> WebSearchTool20250305Param | None:
+        if (
+            tool.name == "web_search"
+            and tool.options
+            and "anthropic" in tool.options
+            and self.model_name in WEB_SEARCH_COMPATIBLE_MODELS
+        ):
+            return _web_search_tool_param(tool.options["anthropic"])
+        else:
+            return None
+
+
+def _web_search_tool_param(
+    maybe_anthropic_options: object,
+) -> WebSearchTool20250305Param:
+    if maybe_anthropic_options is not None and not isinstance(
+        maybe_anthropic_options, dict
+    ):
+        raise TypeError(
+            f"Expected a dictionary for anthropic_options, got {type(maybe_anthropic_options)}"
+        )
+
+    result = WebSearchTool20250305Param(
+        name="web_search",
+        type="web_search_20250305",
+    )
+
+    if maybe_anthropic_options:
+        if "allowed_domains" in maybe_anthropic_options:
+            result["allowed_domains"] = maybe_anthropic_options["allowed_domains"]
+        if "blocked_domains" in maybe_anthropic_options:
+            result["blocked_domains"] = maybe_anthropic_options["blocked_domains"]
+        if "cache_control" in maybe_anthropic_options:
+            result["cache_control"] = maybe_anthropic_options["cache_control"]
+        if "max_uses" in maybe_anthropic_options:
+            result["max_uses"] = maybe_anthropic_options["max_uses"]
+        if "user_location" in maybe_anthropic_options:
+            result["user_location"] = maybe_anthropic_options["user_location"]
+
+    return result
+
 
 # tools can be either a stock tool param or a special Anthropic native use tool param
 ToolParamDef = (
@@ -605,6 +724,8 @@ ToolParamDef = (
     | BetaToolComputerUse20250124Param
     | ToolTextEditor20250124Param
     | BetaToolTextEditor20241022Param
+    | BetaToolTextEditor20250429Param
+    | WebSearchTool20250305Param
 )
 
 
@@ -614,6 +735,8 @@ def add_cache_control(
     | BetaToolComputerUse20250124Param
     | ToolTextEditor20250124Param
     | BetaToolTextEditor20241022Param
+    | BetaToolTextEditor20250429Param
+    | WebSearchTool20250305Param
     | dict[str, Any],
 ) -> None:
     cast(dict[str, Any], param)["cache_control"] = {"type": "ephemeral"}
@@ -698,6 +821,8 @@ async def message_param(message: ChatMessage) -> MessageParam:
                     | ImageBlockParam
                     | ThinkingBlockParam
                     | RedactedThinkingBlockParam
+                    | ServerToolUseBlockParam
+                    | WebSearchToolResultBlockParam
                 ]
             ) = message.error.message
             # anthropic requires that content be populated when
@@ -735,6 +860,8 @@ async def message_param(message: ChatMessage) -> MessageParam:
             | RedactedThinkingBlockParam
             | ImageBlockParam
             | ToolUseBlockParam
+            | ServerToolUseBlockParam
+            | WebSearchToolResultBlockParam
         ] = (
             [TextBlockParam(type="text", text=message.content or NO_CONTENT)]
             if isinstance(message.content, str)
@@ -785,7 +912,7 @@ async def model_output_from_message(
     model: str,
     message: Message,
     tools: list[ToolInfo],
-) -> ModelOutput:
+) -> tuple[ModelOutput, bool]:
     # extract content and tool calls
     content: list[Content] = []
     reasoning_tokens = 0
@@ -800,7 +927,20 @@ async def model_output_from_message(
                 content_text = content_text.replace("<result>", "").replace(
                     "</result>", ""
                 )
-            content.append(ContentText(type="text", text=content_text))
+            content.append(
+                ContentText(
+                    type="text",
+                    text=content_text,
+                    citations=(
+                        [
+                            to_inspect_citation(citation)
+                            for citation in content_block.citations
+                        ]
+                        if content_block.citations
+                        else None
+                    ),
+                )
+            )
         elif isinstance(content_block, ToolUseBlock):
             tool_calls = tool_calls or []
             (tool_name, internal_name) = _names_for_tool_call(content_block.name, tools)
@@ -812,6 +952,10 @@ async def model_output_from_message(
                     internal=internal_name,
                 )
             )
+        elif isinstance(content_block, ServerToolUseBlock):
+            content.append(ContentData(data=content_block.model_dump()))
+        elif isinstance(content_block, WebSearchToolResultBlock):
+            content.append(ContentData(data=content_block.model_dump()))
         elif isinstance(content_block, RedactedThinkingBlock):
             content.append(
                 ContentReasoning(reasoning=content_block.data, redacted=True)
@@ -827,11 +971,12 @@ async def model_output_from_message(
             )
 
     # resolve choice
+    stop_reason, pause_turn = message_stop_reason(message)
     choice = ChatCompletionChoice(
         message=ChatMessageAssistant(
             content=content, tool_calls=tool_calls, model=model, source="generate"
         ),
-        stop_reason=message_stop_reason(message),
+        stop_reason=stop_reason,
     )
 
     # return ModelOutput
@@ -844,17 +989,20 @@ async def model_output_from_message(
         + (input_tokens_cache_read or 0)
         + message.usage.output_tokens  # includes reasoning tokens
     )
-    return ModelOutput(
-        model=message.model,
-        choices=[choice],
-        usage=ModelUsage(
-            input_tokens=message.usage.input_tokens,
-            output_tokens=message.usage.output_tokens,
-            total_tokens=total_tokens,
-            input_tokens_cache_write=input_tokens_cache_write,
-            input_tokens_cache_read=input_tokens_cache_read,
-            reasoning_tokens=reasoning_tokens if reasoning_tokens > 0 else None,
+    return (
+        ModelOutput(
+            model=message.model,
+            choices=[choice],
+            usage=ModelUsage(
+                input_tokens=message.usage.input_tokens,
+                output_tokens=message.usage.output_tokens,
+                total_tokens=total_tokens,
+                input_tokens_cache_write=input_tokens_cache_write,
+                input_tokens_cache_read=input_tokens_cache_read,
+                reasoning_tokens=reasoning_tokens if reasoning_tokens > 0 else None,
+            ),
         ),
+        pause_turn,
     )
 
 
@@ -880,6 +1028,7 @@ def _names_for_tool_call(
         (INTERNAL_COMPUTER_TOOL_NAME, "computer_20250124", "computer"),
         ("str_replace_editor", "text_editor_20241022", "text_editor"),
         ("str_replace_editor", "text_editor_20250124", "text_editor"),
+        ("str_replace_based_edit_tool", "text_editor_20250429", "text_editor"),
         ("bash", "bash_20250124", "bash_session"),
     )
 
@@ -893,16 +1042,18 @@ def _names_for_tool_call(
     )
 
 
-def message_stop_reason(message: Message) -> StopReason:
+def message_stop_reason(message: Message) -> tuple[StopReason, bool]:
     match message.stop_reason:
         case "end_turn" | "stop_sequence":
-            return "stop"
+            return "stop", False
         case "tool_use":
-            return "tool_calls"
+            return "tool_calls", False
         case "max_tokens":
-            return message.stop_reason
+            return message.stop_reason, False
+        case "refusal":
+            return "content_filter", False
         case _:
-            return "unknown"
+            return "unknown", message.stop_reason == "pause_turn"
 
 
 def split_system_messages(
@@ -918,9 +1069,24 @@ def split_system_messages(
 
 async def message_param_content(
     content: Content,
-) -> TextBlockParam | ImageBlockParam | ThinkingBlockParam | RedactedThinkingBlockParam:
+) -> (
+    TextBlockParam
+    | ImageBlockParam
+    | ThinkingBlockParam
+    | RedactedThinkingBlockParam
+    | ServerToolUseBlockParam
+    | WebSearchToolResultBlockParam
+):
     if isinstance(content, ContentText):
-        return TextBlockParam(type="text", text=content.text or NO_CONTENT)
+        citations = (
+            [to_anthropic_citation(citation) for citation in content.citations]
+            if content.citations
+            else None
+        )
+
+        return TextBlockParam(
+            type="text", text=content.text or NO_CONTENT, citations=citations
+        )
     elif isinstance(content, ContentImage):
         # resolve to url
         image = await file_as_data_uri(content.image)
@@ -948,6 +1114,19 @@ async def message_param_content(
             return ThinkingBlockParam(
                 type="thinking", thinking=content.reasoning, signature=content.signature
             )
+    elif isinstance(content, ContentData):
+        match content.data.get("type", None):
+            case "server_tool_use":
+                return cast(
+                    ServerToolUseBlockParam,
+                    ServerToolUseBlock.model_validate(content.data).model_dump(),
+                )
+            case "web_search_tool_result":
+                return cast(
+                    WebSearchToolResultBlockParam,
+                    WebSearchToolResultBlock.model_validate(content.data).model_dump(),
+                )
+        raise NotImplementedError()
     else:
         raise RuntimeError(
             "Anthropic models do not currently support audio or video inputs."
@@ -990,3 +1169,7 @@ def model_call_filter(key: JsonValue | None, value: JsonValue) -> JsonValue:
         value = copy(value)
         value.update(data=BASE_64_DATA_REMOVED)
     return value
+
+
+def _content_list(input: str | list[Content]) -> list[Content]:
+    return [ContentText(text=input)] if isinstance(input, str) else input

inspect_ai/model/_providers/google.py

@@ -26,6 +26,7 @@ from google.genai.types import (
     GenerateContentResponse,
     GenerateContentResponsePromptFeedback,
     GenerateContentResponseUsageMetadata,
+    GoogleSearch,
     HarmBlockThreshold,
     HarmCategory,
     HttpOptions,
@@ -48,6 +49,7 @@ from inspect_ai._util.content import (
 )
 from inspect_ai._util.content import (
     ContentAudio,
+    ContentData,
     ContentImage,
     ContentReasoning,
     ContentText,
@@ -74,6 +76,7 @@ from inspect_ai.model import (
     TopLogprob,
 )
 from inspect_ai.model._model_call import ModelCall
+from inspect_ai.model._providers._google_citations import get_candidate_citations
 from inspect_ai.tool import (
     ToolCall,
     ToolChoice,
@@ -247,7 +250,7 @@ class GoogleGenAIAPI(ModelAPI):
 
         # Create google-genai types.
         gemini_contents = await as_chat_messages(client, input)
-        gemini_tools = chat_tools(tools) if len(tools) > 0 else None
+        gemini_tools = self.chat_tools(tools) if len(tools) > 0 else None
         gemini_tool_config = chat_tool_config(tool_choice) if len(tools) > 0 else None
         parameters = GenerateContentConfig(
             http_options=HttpOptions(headers={HttpHooks.REQUEST_ID_HEADER: request_id}),
@@ -362,6 +365,61 @@ class GoogleGenAIAPI(ModelAPI):
         else:
             return None
 
+    def _use_native_search(self, tool: ToolInfo) -> bool:
+        return (
+            tool.name == "web_search"
+            and tool.options is not None
+            and "gemini" in tool.options
+            # Support "starts with" Gemini 2.0
+            and (self.is_gemini() and not self.is_gemini_1_5())
+        )
+
+    def _categorize_tool(
+        self, acc: tuple[bool, list[FunctionDeclaration]], tool: ToolInfo
+    ) -> tuple[bool, list[FunctionDeclaration]]:
+        """Reducer function that categorizes tools into native search vs function declarations.
+
+        Returns:
+            Tuple of (has_native_search, function_declarations) where has_native_search
+            is True if any tool uses native search, and function_declarations contains
+            all non-native-search tools converted to FunctionDeclaration objects.
+        """
+        return (
+            (True, acc[1])
+            if self._use_native_search(tool)
+            else (
+                acc[0],
+                acc[1]
+                + [
+                    FunctionDeclaration(
+                        name=tool.name,
+                        description=tool.description,
+                        parameters=schema_from_param(tool.parameters)
+                        if len(tool.parameters.properties) > 0
+                        else None,
+                    )
+                ],
+            )
+        )
+
+    def chat_tools(self, tools: list[ToolInfo]) -> ToolListUnion:
+        has_native_search, function_declarations = functools.reduce(
+            self._categorize_tool, tools, (False, list[FunctionDeclaration]())
+        )
+
+        # TODO: Google doesn't yet support native search concurrently with other tools.
+        # Revisit this from time to time to adapt when they fix it.
+        if has_native_search and function_declarations:
+            raise ValueError(
+                "Gemini does not yet support native search concurrently with other tools."
+            )
+
+        return (
+            [Tool(google_search=GoogleSearch())]
+            if has_native_search
+            else [Tool(function_declarations=function_declarations)]
+        )
+
 
 def safety_settings_to_list(
     safety_settings: list[SafetySettingDict],
@@ -500,6 +558,8 @@ async def content_part(client: Client, content: InspectContent | str) -> Part:
         return Part.from_text(text=content.text or NO_CONTENT)
     elif isinstance(content, ContentReasoning):
         return Part.from_text(text=content.reasoning or NO_CONTENT)
+    elif isinstance(content, ContentData):
+        assert False, "Google provider should never encounter ContentData"
     else:
         return await chat_content_to_part(client, content)
 
@@ -538,20 +598,6 @@ async def extract_system_message_as_parts(
     return system_parts or None
 
 
-def chat_tools(tools: list[ToolInfo]) -> ToolListUnion:
-    declarations = [
-        FunctionDeclaration(
-            name=tool.name,
-            description=tool.description,
-            parameters=schema_from_param(tool.parameters)
-            if len(tool.parameters.properties) > 0
-            else None,
-        )
-        for tool in tools
-    ]
-    return [Tool(function_declarations=declarations)]
-
-
 # https://ai.google.dev/gemini-api/tutorials/extract_structured_data#define_the_schema
 def schema_from_param(
     param: ToolParam | ToolParams, nullable: bool | None = False
@@ -656,19 +702,36 @@ def completion_choice_from_candidate(
                 | ContentImage
                 | ContentAudio
                 | ContentVideo
+                | ContentData
             ]
         ) = ""
     # content.parts can be None when the finish_reason is MALFORMED_FUNCTION_CALL
     elif candidate.content.parts is None:
         content = ""
     else:
-        content = []
-        for part in candidate.content.parts:
-            if part.text is not None:
-                if part.thought is True:
-                    content.append(ContentReasoning(reasoning=part.text))
-                else:
-                    content.append(ContentText(text=part.text))
+        # Google's grounded search metadata provides start/end indices for cited
+        # text based on the joining of all separate text parts (despite the doc
+        # suggesting that they provide part_index). Thankfully, the doc also says:
+        #
+        #   Exactly one field within a Part should be set, representing the specific type
+        #   of content being conveyed. Using multiple fields within the same `Part`
+        #   instance is considered invalid.
+        #
+        # That means that we can safely collapse adjacent parts with a `text` field
+        # and not fear that we're breaking other types of content parts
+        parts = functools.reduce(
+            _combine_text_parts, candidate.content.parts, list[Part]()
+        )
+
+        content = [
+            ContentReasoning(reasoning=part.text)
+            if part.thought is True
+            else ContentText(
+                text=part.text, citations=get_candidate_citations(candidate)
+            )
+            for part in parts
+            if part.text is not None
+        ]
 
     # now tool calls
     tool_calls: list[ToolCall] = []
@@ -922,3 +985,16 @@ async def file_for_content(
         files_db.put(content_sha256, str(upload.name))
         # return the file
         return upload
+
+
+def _combine_text_parts(acc: list[Part], part: Part) -> list[Part]:
+    """Combine adjacent text parts into a single part."""
+    return (
+        acc + [part]
+        if part.text is None
+        or part.thought is True
+        or len(acc) == 0
+        or acc[-1].text is None
+        or acc[-1].thought is True
+        else acc[:-1] + [Part(text=acc[-1].text + part.text)]
+    )

inspect_ai/model/_providers/mistral.py

@@ -44,9 +44,15 @@ from typing_extensions import override
 # TODO: Migration guide:
 # https://github.com/mistralai/client-python/blob/main/MIGRATION.md
 from inspect_ai._util.constants import NO_CONTENT
-from inspect_ai._util.content import Content, ContentImage, ContentText
+from inspect_ai._util.content import (
+    Content,
+    ContentImage,
+    ContentReasoning,
+    ContentText,
+)
 from inspect_ai._util.http import is_retryable_http_status
 from inspect_ai._util.images import file_as_data_uri
+from inspect_ai.model._reasoning import parse_content_with_reasoning
 from inspect_ai.tool import ToolCall, ToolChoice, ToolFunction, ToolInfo
 
 from ..._util.httpx import httpx_should_retry
@@ -481,26 +487,33 @@ def completion_content(content: str | list[ContentChunk]) -> str | list[Content]
     if isinstance(content, str):
         return content
     else:
-        return [completion_content_chunk(c) for c in content]
+        return [item for c in content for item in completion_content_chunks(c)]
 
 
-def completion_content_chunk(content: ContentChunk) -> Content:
+def completion_content_chunks(content: ContentChunk) -> list[Content]:
     if isinstance(content, ReferenceChunk):
         raise TypeError("ReferenceChunk content is not supported by Inspect.")
     elif isinstance(content, TextChunk):
-        return ContentText(text=content.text)
+        parsed = parse_content_with_reasoning(content.text)
+        if parsed:
+            return [
+                ContentReasoning(reasoning=parsed.reasoning),
+                ContentText(text=parsed.content),
+            ]
+        else:
+            return [ContentText(text=content.text)]
     elif isinstance(content, DocumentURLChunk):
-        return ContentText(text=content.document_url)
+        return [ContentText(text=content.document_url)]
     else:
         if isinstance(content.image_url, str):
-            return ContentImage(image=content.image_url)
+            return [ContentImage(image=content.image_url)]
         else:
             match content.image_url.detail:
                 case "low" | "high":
                     detail: Literal["auto", "low", "high"] = content.image_url.detail
                 case _:
                     detail = "auto"
-            return ContentImage(image=content.image_url.url, detail=detail)
+            return [ContentImage(image=content.image_url.url, detail=detail)]
 
 
 def completion_choices_from_response(